We carefully reviewed the study published by Sharma et al. [1] in the Clinical and Molecular Hepatology, which combined plasma lipidomics, fungal peptidomics, and machine learning methods to predict the risk of early death in patients with acute liver failure (ALF). However, the study has some methodological and reporting issues that warrant further discussion.

First, in the Methods, the authors define non-survivors as patients who die within 30 days and survivors as those who remain alive for at least 90 days. However, in the results, they report an overall 90-day mortality of about 65% (178/270) and plot 30-day Kaplan-Meier survival curves according to a probability of detection (POD)-lipid threshold. This leaves it unclear how patients who die between 30 and 90 days and those who undergo liver transplantation were counted, whether as deaths, survivors, or a separate competing outcome. This is problematic because, as Stravitz and Lee point out, acute liver failure studies are expected to state follow-up windows and the handling of transplantation explicitly [2]. Furthermore, in the multivariate Cox model, the hazard ratio for POD lipids was 1.99 (95% confidence interval, 1.02–2.04). However, recalculating the estimate on a logarithmic scale yields approximately 1.4, not 1.99, indicating an error in the reported hazard ratio. If the true hazard ratio is 1.4, POD lipids would still be an independent predictor of early death, but the increased risk would be approximately 40%, not 99%. This weakens its strength as a superior prognostic marker compared to traditional scores.

Second, because this study analyzed thousands of lipids, fungal peptides, and clinical variables, it is crucial to report in detail the methods used to handle multiple comparisons. In the study, the Methods section states that P-values were adjusted using the Benjamini-Hochberg method, controlling the false discovery rate at <0.01. However, the volcano plots, module-trait heatmaps, and clinical correlation matrices show nominal P<0.05, without indicating which values were adjusted or providing q-values. This makes it difficult to determine how many associations remain statistically significant after correction. Previous research by Trovato et al. [3] has already shown that in this high-dimensional data context, the associations between lipids and mortality in ALF patients should be carefully analyzed and reported. Additionally, multiple receiver operating characteristic curves for biomarker combinations containing five lipids or six peptides reported area under the curve values of 1.00 in this exploratory study cohort, indicating that these biomarker combinations could effectively distinguish survivors from non-survivors in a small sample. However, this may be due to overfitting, exaggerating the actual predictive value of these biomarker combinations, and therefore requires confirmation in a completely independent validation cohort.

Third, the researchers used a simple linear regression model to analyze the relationship between the Shannon diversity index and binary complications such as infection, necrosis, and hepatic encephalopathy. Previous studies of acute liver failure cohorts typically used logistic regression or time-to-event models to analyze such outcomes, adjusting for baseline severity (e.g., model for end-stage liver disease score, King’s College Hospital criteria), a method that better reflects the outcome structure and reduces residual confounding factors [2]. In the machine learning phase, the researchers selected the top five lipids in an exploratory cohort of 40 patients, but then trained and evaluated the artificial neural network on all 270 ALF cases using a single 70/30 internal split, without using a completely independent test set. Unlike the more robust procedures in previous lipidomics prognostic studies of ALF,3 this design may lead to an overly optimistic estimate of accuracy in this single-center study.

The graphical abstract and discussion section further suggest that downregulation of Clec7a “leads to fungal infection” and promotes the accumulation of phosphatidylcholine and phosphatidic acid through the “Kennedy pathway.” However, the data provided mainly consist of cross-sectional plasma analysis results and lack host transcriptome data, functional assays, or data on confirmed invasive fungal infections. Therefore, these causal claims are largely speculative, and greater caution is needed when interpreting the study results and generalizing the conclusions.

Citations to this article as recorded by